Material map generation using hyper-spectral NIR images

Dong-Keun Han, Jeong-Won Ha, Jong-Ok Kim

Abstract: The hyper-spectral curve on the near-infrared (NIR) bands commonly exhibits distinct characteristics for each surface material. NIR information can be a useful clue to identify the surface material of an object. In this paper, the surface material of each local patch is first classified by a deep network from NIR hyper-spectral images, and then, those classification results are collected to obtain the surface material map of an entire scene. To train the classification network, we built a hyper-spectral dataset which includes 5 different materials. Experimental results show that we can get a quite effective material map.